Gemini vs. OpenAI: will Google bleed out the competition or face its own innovator's dilemma?

Summary

Google's pricing discipline and the monopoly myth

Ross Hendrix argues that if Gemini wins the AI race, Google could keep the model free until rivals collapse, then monetize a monopoly. The argument relies on Google's cash reserves and willingness to wage a price war. Google does have the scale to do this. It already spends billions outside OpenAI's core business on consumer electronics, science, and chips.

But Google is a public company. Investors care about near-term margins. When you own the most profitable business in tech—search with 80% historical margins—a free product that doesn't monetize immediately invites Wall Street skepticism. Meta faced this tension with Reels. Instagram's shift to short-form video looked like a margin-destroying move until the company proved it could monetize video at comparable rates. Google would face the same shareholder pressure with Gemini.

Both Gemini and ChatGPT now charge comparable prices for premium tiers. Neither is running at a loss. The LLM market is unlikely to collapse into a single monopoly. A duopoly seems more probable at minimum, with Anthropic well-positioned. Dario Amodei's recent comments at Deal Book suggested confidence in Anthropic's position.

AMD's silicon defense

Lisa Su broke silence on Google's TPU in a move that reads as AMD protecting its GPU narrative. Her argument is familiar: TPUs are purpose-built for specific workloads, GPUs offer flexibility. But empirical backing matters. Researchers doing undirected foundation model work might actually prefer GPUs. It's unclear why a researcher would choose TPUs over GPUs if flexibility mattered most.

Su's underlying claim is that ASICs like TPUs will capture 20 to 25 percent of the accelerator market over five years while GPUs hold the majority. This seems defensible. At companies like Meta, transformer-based LLM inference accounts for less than 20 percent of compute spend. The rest goes to recommendation systems, content delivery, and ad placement, workloads that don't benefit from transformer-specific hardware. AMD and Nvidia both know this. Nvidia argues it offers greater performance, versatility, and fungibility than ASICs, essentially selling a toolbox. AMD could replicate what Nvidia has already done: create specialized chips for specific foundation model companies and pitch them the upside of a custom solution. Nvidia forked gaming and AI GPUs years ago.

Ads in ChatGPT: waiting for the first domino

OpenAI has reportedly pulled back on ad rollout plans, but the real story is timing. Sean Frank, founder of Ridge, reported that ChatGPT referrals convert at 12 percent, roughly ten times better than typical ecommerce conversion rates of 1 to 2 percent. The intent signal from ChatGPT users is extraordinarily strong. They have already done product research and are ready to buy.

That conversion gap is precisely why both Google and OpenAI will eventually run ads. The question is who moves first. Ben Thompson, Eric Soufford, and others have signaled support for monetization, but the first ad in a chat app will be screenshotted and shared globally. Neither company wants that headline. Google might let OpenAI take the hit first, a rational game of chicken where OpenAI, under capital pressure, breaks first.

Google could run ads at a loss to force competitors into monetizing, but Google has historically resisted price wars. It didn't wage one against AWS and Azure, instead competing on functionality and integration. The ad infrastructure and customer relationships already exist at Google. Running loss-leader ads costs less there than at OpenAI.

Sam Altman's rocket bet

OpenAI's CEO explored acquiring or partnering with a rocket company to compete with Elon Musk's SpaceX. Discussions with Stokes Space, valued at $2 billion as of October, picked up in fall but are no longer active. The Journal reports Altman wanted a controlling stake and was willing to invest billions over time.

This is OpenAI's tenth competitive front. The company is burning capital on multiple bets simultaneously across space infrastructure, robotics, and consumer hardware before achieving profitability. Uber and Amazon both burned heavily before becoming cash-flow positive, but their burn was measured in billions, not the $140 billion OpenAI is projected to consume before turning a profit, per a Deutsche Bank chart cited by Joe Weisenthal. Investors worry Altman is trying to own the full stack: chips, data centers, rockets, and eventually the sand to make the chips. Bezos, who has Blue Origin and Rivian but lacks a major AI bet, would seem a natural ally, yet no partnership emerged.

Space data centers solve a political problem. They remove the local electricity bill and water use complaints that plague terrestrial facilities. But skeptics note that every time space data center concepts go viral, people dismiss them as physics-defying fantasies, even as many players quietly explore them. The real barrier is capital and regulatory approval, not technological implausibility.

The unsigned $100 billion deal

Nvidia's $100 billion OpenAI mega-deal remains unsigned two months after its announcement. Nvidia disclosed in earnings that it's in letter of intent phase, not fully executed. This likely reflects the economics of how Nvidia discounts chips in exchange for equity upside. Dylan Patel calculated the effective discount at roughly 30 percent. Jensen Huang's appearance on Joe Rogan shortly after suggested a defensive posture. He's fighting to keep OpenAI on Nvidia GPUs while Google pushes TPUs and OpenAI openly explores alternatives.

The deal will probably close in some form. But Huang can't afford to signal that Nvidia is betting its future on a single customer.